DeepSeek Adds Vision Capabilities to v4-flash Experimental Model
DeepSeek has introduced vision capabilities to its v4-flash experimental model, allowing users to input images directly via the API. The model automatically resizes these images and bills them as tokens alongside text inputs. This update provides developers with a more cost-effective, multimodal option for tasks requiring both text and visual processing, reducing the need to switch to alternative models for basic image tasks. Before inference, images are automatically resized: those below 384x384 pixels are scaled up, while larger images are scaled down to roughly 800x800 pixels to preserve the aspect ratio before being converted into tokens.
## BACKGROUND
Multimodal Large Language Models (MLLMs) process images by converting them into visual tokens, mapping pixels into high-dimensional vectors in an embedding space. This allows the underlying language model to process visual data in the same format as text tokens.